AMD Reportedly In Advanced Talks To Buy Xilinx for Roughly $30 Billion

[coppice]

I'm really not a fan of all this consolidation. Every time it happens a lot of parts get discontinued, a lot of people lose their jobs, and we are left with fewer choices and less competition. Maybe it's just part of the inevitable race to the bottom that is going to leave us with the Chinese companies owning the entire market.

The pace of consolidation in the electronics market has been frightening the past 5 or 10 years. While I think AMD would be a stronger company with its own FPGA branch, I don't think that condensing a diverse field of respected companies into a few behemoths serves anyone. It doesn't tend to be good for either company for various reasons, nor the customer or state of technology. I guess only the shareholders win. I can't help but feel that the free market and big money inevitably lead to a perverted version of a good thing with people losing out all round.

You are going to have behemoths somewhere along the supply chain. The real consolidation has been in fabs, not component vendors. Fabs now need to be so big and expensive to be viable, that there are very few players there, and its hard to see how it could be otherwise. We still have numerous choices of supplier for most semiconductor devices, but the device vendors have very few choices when it comes to fabbing their silicon. When a company like TSMC has to close an old process, as no longer economically viable, its quite possible that every device of a certain type, from every device vendor, has to be withdrawn at the same time, as they were all made in the same fab.

[tszaboo]

I've never dealt with AMD before so I don't know what they're like. I had an Athlon PC around 20 years ago that was fantastic but then it seems like they didn't have much to offer for a while and I've been on Intel ever since. AMD was nearly irrelevant for a while and I thought they might go under, glad to hear they're growing though because Intel needs some competition.

I heared a story, about dealing with them. The company wanted a linux SBC, that time there was AMD geode. So they introduced the new selected part, then managed to obsolete the entire series, before they could fully assemble the first prototype batch. The parts were quite competitive and made sense back in ~2000.
Same with intel and their curie platform. They were targeting the Arduino folks, and microcontrollers with the product. It had like 512 MB of RAM or some other silliness only BGA package, and their datasheet was kept secret. Like, I dont even understand, how they could miss their target audience. From the same company that made the 8051, and sold billions of them. They released something that was not even in the ballpark of being usable.

[brucehoult]

When a company like TSMC has to close an old process, as no longer economically viable, its quite possible that every device of a certain type, from every device vendor, has to be withdrawn at the same time, as they were all made in the same fab.

Which is not going to happen if there are customers using the fab and willing to pay about the same for the dies as they always were, or a bit less as the factory is all paid off and depreciated.

If the equipment in it starts to fail and need a lot of expensive repairs that's a different thing, but I imagine it has a longer lifetime than people want to keep using an old process anyway.

[David Hess]

Given that Intel made that exact move with Altera a while ago, this is yet another simple reason why AMD would do just the same. AMD is a direct Intel competitor, and as such, they have little choice but expand their offering in the same way, otherwise they're bound to lose ground one way or another in the long run. This is pretty common in the industry. Major competitors tend to expand their offerings in the same way. Just my 2 cents.

Intel bought Altera as part of their effort to fill their foundries and provide foundry services.  AMD has no such incentive to do the same with Xilinx since they divested their foundries to GlobalFoundries and now use TSMC.

If Intel had a plan for CPU compatible FPGAs, it has not worked out.

[coppice]

When a company like TSMC has to close an old process, as no longer economically viable, its quite possible that every device of a certain type, from every device vendor, has to be withdrawn at the same time, as they were all made in the same fab.

Which is not going to happen if there are customers using the fab and willing to pay about the same for the dies as they always were, or a bit less as the factory is all paid off and depreciated.

If the equipment in it starts to fail and need a lot of expensive repairs that's a different thing, but I imagine it has a longer lifetime than people want to keep using an old process anyway.

Some people still want to be able to buy parts, especially analogue parts, that were launched 40 years ago. When a fab reaches its end of life, its seldom a sudden catastrophe. Most vendors are pretty good about planning for a shut down, notifying customers, and fulfilling lifetime buys of parts before the fab closes. If you don't order in time, that's your problem.

You must think TSMC is a magic company, immune from the pressures the rest of the industry faces. A lot of the mass removals of parts from vendor's catalogues that have occurred over the years have been due to a fab getting so old and creaky, the only production facility for the part had to close. You do know that TSMC is no longer running all the processes it has had since day one, right?

[David Hess]

Some people still want to be able to buy parts, especially analogue parts, that were launched 40 years ago. When a fab reaches its end of life, its seldom a sudden catastrophe. Most vendors are pretty good about planning for a shut down, notifying customers, and fulfilling lifetime buys of parts before the fab closes. If you don't order in time, that's your problem.

Fabrication facilities may also be closed because the raw materials becomes unavailable.

You must think TSMC is a magic company, immune from the pressures the rest of the industry faces. A lot of the mass removals of parts from vendor's catalogues that have occurred over the years have been due to a fab getting so old and creaky, the only production facility for the part had to close. You do know that TSMC is no longer running all the processes it has had since day one, right?

Unlike Intel, TSMC maintains several generations of older processes.

[brucehoult]

You must think TSMC is a magic company, immune from the pressures the rest of the industry faces. A lot of the mass removals of parts from vendor's catalogues that have occurred over the years have been due to a fab getting so old and creaky, the only production facility for the part had to close. You do know that TSMC is no longer running all the processes it has had since day one, right?

Ahahaha. Of course no.

But they're still doing 180, which debuted in 1998 I think.
And even 350 nm from ... what? 1994 or so? ... is still available.

I think 500 nm has been gone for a while though. And both the above are I think only on older 200mm wafers.

TSMC only started in 1987.

[Berni]

Its mostly digital fab processes that are marching on at such a quick pace.

The analog fab processes have much longer lifetimes since they don't really benefit from miniaturization as much since the transistor counts are much lower and dies are small anyway. So it makes sense to keep a lot of the older fabs running. However when a fab does get shut down its much harder to move to a new one since analog chips might use some special fine tuned steps to get the analog performance required. While transitioning a simple digital IC to a new process is a lot easier (Tho there are cases where the new process made the chips faster than the original, so this made some of the products using them not work anymore as they needed the chip to be slow)

Going back to AMD. They are really kicking ass recently tho. The just released Zen 3 pretty much put the final nail in Intels coffin as AMD has caught up on the only thing that Intel still had an advantage in, the performance per cycle metric. Finally Intel has some motivation to actually do something ever since the 4th Gen i5/i7.

[coppice]

But they're still doing 180, which debuted in 1998 I think.
And even 350 nm from ... what? 1994 or so? ... is still available.

I think 500 nm has been gone for a while though. And both the above are I think only on older 200mm wafers.

I suspect 350nm will hang around for a long time. It has an interesting and useful position. Its the last step before the operating voltages get really low. 350nm devices typically run up to 3.6V throughout the chip. 250nm devices are usually limited to 2.5V, requiring split supplies and low density higher voltage transistors around the perimeter of the chip to be able to interface to a higher voltage external world. This means 250nm is not long for this world. Once you have to split the core and peripheral voltages you are usually better off going for a finer geometry. However, wafer costs for 350nm are now rising over time, so there will be an increasing detterence factor.

[SMB784]

But they're still doing 180, which debuted in 1998 I think.
And even 350 nm from ... what? 1994 or so? ... is still available.

I think 500 nm has been gone for a while though. And both the above are I think only on older 200mm wafers.

I suspect 350nm will hang around for a long time. It has an interesting and useful position. Its the last step before the operating voltages get really low. 350nm devices typically run up to 3.6V throughout the chip. 250nm devices are usually limited to 2.5V, requiring split supplies and low density higher voltage transistors around the perimeter of the chip to be able to interface to a higher voltage external world. This means 250nm is not long for this world. Once you have to split the core and peripheral voltages you are usually better off going for a finer geometry. However, wafer costs for 350nm are now rising over time, so there will be an increasing detterence factor.

Interesting observation. I guess my question is why is running things at 3.6V preferable than say 2.5 or 1.5 if you can get away with it?

[coppice]

I guess my question is why is running things at 3.6V preferable than say 2.5 or 1.5 if you can get away with it?

Low voltages outside the silicon can bring a lot of issues. They are normal in some situations, like PC motherboards, now. However, inside electrically noisy equipment, like household electrical appliances, they can be very troublesome. Deep in the heart of an IC they are usually fine, but if the chip needs split rails to allow for higher interface voltages, it becomes more complex. This overhead is no issue for complex ICs, which usually gain more from a fine geometry than they lose. The increased leakage in fine geometry processes can be an issue in ultra low power applications, but that's the main problem area for complex ICs. For simpler ICs the overhead due to split rail operation can be serious.

[ali_asadzadeh]

This would make the game very easy for Chinese FPGA manufacturers, they will dominate very soon.

[KaneTW]

For anyone that's not a giant, AMD will just tell you to sod off. I have/had a project based on x86 CPUs. AMD referred me to SBCs (which are not suitable at all), despite this project having 6 digit funding. I guess they don't care about you until you're in the 7-8 digits.

[coppice]

This would make the game very easy for Chinese FPGA manufacturers, they will dominate very soon.

This won't affect the existing Chinese FPGA manufacturers much. They are playing in a different, low end, market. What will be affecting FPGA development in China is the big 5G makers being cut off from American FPGAs. That must be spurring a lot of high end FPGA development in China right now.

[ralphrmartin]

When a company like TSMC has to close an old process, as no longer economically viable, its quite possible that every device of a certain type, from every device vendor, has to be withdrawn at the same time, as they were all made in the same fab.

Which is not going to happen if there are customers using the fab and willing to pay about the same for the dies as they always were, or a bit less as the factory is all paid off and depreciated.

It may not be as simple as that. They may need the space the line occupies for a new line, for example.

[SilverSolder]

This would make the game very easy for Chinese FPGA manufacturers, they will dominate very soon.

This won't affect the existing Chinese FPGA manufacturers much. They are playing in a different, low end, market. What will be affecting FPGA development in China is the big 5G makers being cut off from American FPGAs. That must be spurring a lot of high end FPGA development in China right now.

...which will eventually lead to competition in that space.  Own goal by the USA in my view, but what do I know...

[asmi]

...which will eventually lead to competition in that space.  Own goal by the USA in my view, but what do I know...

That's exactly how ITAR-free satellite market came to be. One would think people can learn at least on their own mistakes.

[SiliconWizard]

Given that Intel made that exact move with Altera a while ago, this is yet another simple reason why AMD would do just the same. AMD is a direct Intel competitor, and as such, they have little choice but expand their offering in the same way, otherwise they're bound to lose ground one way or another in the long run. This is pretty common in the industry. Major competitors tend to expand their offerings in the same way. Just my 2 cents.

Intel bought Altera as part of their effort to fill their foundries and provide foundry services.  AMD has no such incentive to do the same with Xilinx since they divested their foundries to GlobalFoundries and now use TSMC.

If Intel had a plan for CPU compatible FPGAs, it has not worked out.

I think you are strictly seeing it from a technical POV. And you seem to consider Intel strictly as a CPU vendor (with foundry services) not willing or able to expand their horizons. Intel already does more than just design and sell CPUS. Filling their foundries may have been an incentive, but the end result is much more than just that.

From a business POV, buying Altera means Intel is now also an FPGA vendor (with a significant market). The fact they would mix FPGA and CPU technologies could be interesting, but is almost irrelevant at this point. That means, at this point, that Intel has a bigger market share overall as a company. All big companies need to expand in that way in order to survive long term. Diversification is key.

Now regarding mixing technologies, that's probably not for tomorrow, but there could be some interesting opportunities. Intel is not ARM. I don't see - and don't think that would make sense - them, for instance, embedding one of their CPU cores in FPGAs, the way it has already happened with ARM cores. I doubt that would fit Intel's strategy or model. But I could imagine something similar, but IMO more appropriate for Intel's strategy: embedding some tightly-coupled FPGA in their future CPUs, to be used as reconfigurable accelerators for instance. The difference may seem subtle, except that this way they would sell powerful CPUs with some integrated FPGA fabric, and not FPGAs with some integrated and only modest CPU core.

[wraper]

But I could imagine something similar, but IMO more appropriate for Intel's strategy: embedding some tightly-coupled FPGA in their future CPUs, to be used as reconfigurable accelerators for instance. The difference may seem subtle, except that this way they would sell powerful CPUs with some integrated FPGA fabric, and not FPGAs with some integrated and only modest CPU core.

They already do that https://en.wikichip.org/wiki/intel/xeon_gold/6138p#:~:text=Xeon%20Gold%206138P%20is%20a,by%20Intel%20in%20early%2D2018.&text=This%20microprocessor%2C%20which%20operates%20at,channel%20DDR4%2D2666%20ECC%20memory.

[Dmeads]

https://seekingalpha.com/article/4378735-amd-and-xilinx-prize-is-versal-acap-not-fpgas

found this article: really good read.

Basically it says that AMD's real motivation for the 30Billion comes from Xilinx'x ACAP. The article states that the Xilinx core FPGA market has only risen 4.8% since 2013, which isnt worth the 30 billion dolllars.

It also says, despite the fact that FPGAs will replace GPUs for AI acceleration, "that market is still small but with strong growth potential," also not making the deal worth it to AMD (in terms of hardware acceleration).

The ACAP however is targeted towards the fastest growing and "next generation" semiconductor markets such as 5G, defense, autonomous driving assist, etc, which is why the article says AMD wants Xilinx.

The author guesses that Xilinx will reject the offer, due to AMD not having much to offer them.

Any thoughts?

[helius]

It simply looks like a more advanced SoC FPGA, like the Zynq from 10 years ago. Hardened I/O blocks is not new, although the "AI" vector computation units have not been available in FPGAs before (unless I missed something, which is likely).

[SilverSolder]

It simply looks like a more advanced SoC FPGA, like the Zynq from 10 years ago. Hardened I/O blocks is not new, although the "AI" vector computation units have not been available in FPGAs before (unless I missed something, which is likely).

ISTR something about Tesla using this kind of technology, in a custom chip?

[asmi]

Growing popularity of Nvidia's datacenter products tells me you don't need FPGA to accelerate AI/ML. AMD makes their own datacenter-related products too (based on GPU tech), so if anything I'd expect to see FPGA fabric integrated into datacenter "GPUs" rather than into CPUs.

[SiliconWizard]

But I could imagine something similar, but IMO more appropriate for Intel's strategy: embedding some tightly-coupled FPGA in their future CPUs, to be used as reconfigurable accelerators for instance. The difference may seem subtle, except that this way they would sell powerful CPUs with some integrated FPGA fabric, and not FPGAs with some integrated and only modest CPU core.

They already do that https://en.wikichip.org/wiki/intel/xeon_gold/6138p#:~:text=Xeon%20Gold%206138P%20is%20a,by%20Intel%20in%20early%2D2018.&text=This%20microprocessor%2C%20which%20operates%20at,channel%20DDR4%2D2666%20ECC%20memory.

I didn't know about this device yet. That confirms Intel's interest. This just looks like a first step though IMO: it's apparently just an existing FPGA die coupled to a CPU die. I expect something more integrated in the future, like FPGA fabric possibly directly on the same die and possibly more "tightly coupled".

[asmi]

I didn't know about this device yet. That confirms Intel's interest. This just looks like a first step though IMO: it's apparently just an existing FPGA die coupled to a CPU die. I expect something more integrated in the future, like FPGA fabric possibly directly on the same die and possibly more "tightly coupled".

It's more efficient to use separate dies on interposer, because you are not limited to using the same process for all subcomponents, and yields are generally better for smaller dies. And AMD's success in recent years has proven that SiP designs (which is what all their modern CPUs are) are a way to go.